• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright 2019 Google Inc.
3  *
4  * Use of this source code is governed by a BSD-style license that can be
5  * found in the LICENSE file.
6  */
7 
8 #ifndef SKVX_DEFINED
9 #define SKVX_DEFINED
10 
11 // skvx::Vec<N,T> are SIMD vectors of N T's, a v1.5 successor to SkNx<N,T>.
12 //
13 // This time we're leaning a bit less on platform-specific intrinsics and a bit
14 // more on Clang/GCC vector extensions, but still keeping the option open to
15 // drop in platform-specific intrinsics, actually more easily than before.
16 //
17 // We've also fixed a few of the caveats that used to make SkNx awkward to work
18 // with across translation units.  skvx::Vec<N,T> always has N*sizeof(T) size
19 // and alignment and is safe to use across translation units freely.
20 // (Ideally we'd only align to T, but that tanks ARMv7 NEON codegen.)
21 
22 // Please try to keep this file independent of Skia headers.
23 #include <algorithm>         // std::min, std::max
24 #include <cassert>           // assert()
25 #include <cmath>             // ceilf, floorf, truncf, roundf, sqrtf, etc.
26 #include <cstdint>           // intXX_t
27 #include <cstring>           // memcpy()
28 #include <initializer_list>  // std::initializer_list
29 #include <utility>           // std::index_sequence
30 
31 #if defined(__SSE__) || defined(__AVX__) || defined(__AVX2__)
32     #include <immintrin.h>
33 #elif defined(__ARM_NEON)
34     #include <arm_neon.h>
35 #elif defined(__wasm_simd128__)
36     #include <wasm_simd128.h>
37 #endif
38 
39 // To avoid ODR violations, all methods must be force-inlined...
40 #if defined(_MSC_VER)
41     #define SKVX_ALWAYS_INLINE __forceinline
42 #else
43     #define SKVX_ALWAYS_INLINE __attribute__((always_inline))
44 #endif
45 
46 // ... and all standalone functions must be static.  Please use these helpers:
47 #define SI    static inline
48 #define SIT   template <       typename T> SI
49 #define SIN   template <int N            > SI
50 #define SINT  template <int N, typename T> SI
51 #define SINTU template <int N, typename T, typename U, \
52                         typename=std::enable_if_t<std::is_convertible<U,T>::value>> SI
53 
54 namespace skvx {
55 
56 template <int N, typename T>
57 struct alignas(N*sizeof(T)) Vec;
58 
59 template <int... Ix, int N, typename T>
60 SI Vec<sizeof...(Ix),T> shuffle(const Vec<N,T>&);
61 
62 template <typename D, typename S>
63 SI D bit_pun(const S&);
64 
65 // All Vec have the same simple memory layout, the same as `T vec[N]`.
66 template <int N, typename T>
67 struct alignas(N*sizeof(T)) VecStorage {
68     SKVX_ALWAYS_INLINE VecStorage() = default;
VecStorageVecStorage69     SKVX_ALWAYS_INLINE VecStorage(T s) : lo(s), hi(s) {}
70 
71     Vec<N/2,T> lo, hi;
72 };
73 
74 template <typename T>
75 struct VecStorage<4,T> {
76     SKVX_ALWAYS_INLINE VecStorage() = default;
77     SKVX_ALWAYS_INLINE VecStorage(T s) : lo(s), hi(s) {}
78     SKVX_ALWAYS_INLINE VecStorage(T x, T y, T z, T w) : lo(x,y), hi(z, w) {}
79     SKVX_ALWAYS_INLINE VecStorage(Vec<2,T> xy, T z, T w) : lo(xy), hi(z,w) {}
80     SKVX_ALWAYS_INLINE VecStorage(T x, T y, Vec<2,T> zw) : lo(x,y), hi(zw) {}
81     SKVX_ALWAYS_INLINE VecStorage(Vec<2,T> xy, Vec<2,T> zw) : lo(xy), hi(zw) {}
82 
83     SKVX_ALWAYS_INLINE Vec<2,T>& xy() { return lo; }
84     SKVX_ALWAYS_INLINE Vec<2,T>& zw() { return hi; }
85     SKVX_ALWAYS_INLINE T& x() { return lo.lo.val; }
86     SKVX_ALWAYS_INLINE T& y() { return lo.hi.val; }
87     SKVX_ALWAYS_INLINE T& z() { return hi.lo.val; }
88     SKVX_ALWAYS_INLINE T& w() { return hi.hi.val; }
89 
90     SKVX_ALWAYS_INLINE Vec<2,T> xy() const { return lo; }
91     SKVX_ALWAYS_INLINE Vec<2,T> zw() const { return hi; }
92     SKVX_ALWAYS_INLINE T x() const { return lo.lo.val; }
93     SKVX_ALWAYS_INLINE T y() const { return lo.hi.val; }
94     SKVX_ALWAYS_INLINE T z() const { return hi.lo.val; }
95     SKVX_ALWAYS_INLINE T w() const { return hi.hi.val; }
96 
97     // Exchange-based swizzles. These should take 1 cycle on NEON and 3 (pipelined) cycles on SSE.
98     SKVX_ALWAYS_INLINE Vec<4,T> yxwz() const { return shuffle<1,0,3,2>(bit_pun<Vec<4,T>>(*this)); }
99     SKVX_ALWAYS_INLINE Vec<4,T> zwxy() const { return shuffle<2,3,0,1>(bit_pun<Vec<4,T>>(*this)); }
100 
101     Vec<2,T> lo, hi;
102 };
103 
104 template <typename T>
105 struct VecStorage<2,T> {
106     SKVX_ALWAYS_INLINE VecStorage() = default;
107     SKVX_ALWAYS_INLINE VecStorage(T s) : lo(s), hi(s) {}
108     SKVX_ALWAYS_INLINE VecStorage(T x, T y) : lo(x), hi(y) {}
109 
110     SKVX_ALWAYS_INLINE T& x() { return lo.val; }
111     SKVX_ALWAYS_INLINE T& y() { return hi.val; }
112 
113     SKVX_ALWAYS_INLINE T x() const { return lo.val; }
114     SKVX_ALWAYS_INLINE T y() const { return hi.val; }
115 
116     // This exchange-based swizzle should take 1 cycle on NEON and 3 (pipelined) cycles on SSE.
117     SKVX_ALWAYS_INLINE Vec<2,T> yx() const { return shuffle<1,0>(bit_pun<Vec<2,T>>(*this)); }
118 
119     SKVX_ALWAYS_INLINE Vec<4,T> xyxy() const {
120         return Vec<4,T>(bit_pun<Vec<2,T>>(*this), bit_pun<Vec<2,T>>(*this));
121     }
122 
123     Vec<1,T> lo, hi;
124 };
125 
126 template <int N, typename T>
127 struct alignas(N*sizeof(T)) Vec : public VecStorage<N,T> {
128     static_assert((N & (N-1)) == 0,        "N must be a power of 2.");
129     static_assert(sizeof(T) >= alignof(T), "What kind of unusual T is this?");
130 
131     // Methods belong here in the class declaration of Vec only if:
132     //   - they must be here, like constructors or operator[];
133     //   - they'll definitely never want a specialized implementation.
134     // Other operations on Vec should be defined outside the type.
135 
136     SKVX_ALWAYS_INLINE Vec() = default;
137 
138     using VecStorage<N,T>::VecStorage;
139 
140     SKVX_ALWAYS_INLINE Vec(std::initializer_list<T> xs) {
141         T vals[N] = {0};
142         memcpy(vals, xs.begin(), std::min(xs.size(), (size_t)N)*sizeof(T));
143 
144         this->lo = Vec<N/2,T>::Load(vals +   0);
145         this->hi = Vec<N/2,T>::Load(vals + N/2);
146     }
147 
148     SKVX_ALWAYS_INLINE T  operator[](int i) const { return i<N/2 ? this->lo[i] : this->hi[i-N/2]; }
149     SKVX_ALWAYS_INLINE T& operator[](int i)       { return i<N/2 ? this->lo[i] : this->hi[i-N/2]; }
150 
151     SKVX_ALWAYS_INLINE static Vec Load(const void* ptr) {
152         Vec v;
153         memcpy(&v, ptr, sizeof(Vec));
154         return v;
155     }
156     SKVX_ALWAYS_INLINE void store(void* ptr) const {
157         memcpy(ptr, this, sizeof(Vec));
158     }
159 };
160 
161 template <typename T>
162 struct Vec<1,T> {
163     T val;
164 
165     SKVX_ALWAYS_INLINE Vec() = default;
166 
167     Vec(T s) : val(s) {}
168 
169     SKVX_ALWAYS_INLINE Vec(std::initializer_list<T> xs) : val(xs.size() ? *xs.begin() : 0) {}
170 
171     SKVX_ALWAYS_INLINE T  operator[](int) const { return val; }
172     SKVX_ALWAYS_INLINE T& operator[](int)       { return val; }
173 
174     SKVX_ALWAYS_INLINE static Vec Load(const void* ptr) {
175         Vec v;
176         memcpy(&v, ptr, sizeof(Vec));
177         return v;
178     }
179     SKVX_ALWAYS_INLINE void store(void* ptr) const {
180         memcpy(ptr, this, sizeof(Vec));
181     }
182 };
183 
184 // Ideally we'd only use bit_pun(), but until this file is always built as C++17 with constexpr if,
185 // we'll sometimes find need to use unchecked_bit_pun().  Please do check the call sites yourself!
186 template <typename D, typename S>
187 SI D unchecked_bit_pun(const S& s) {
188     D d;
189     memcpy(&d, &s, sizeof(D));
190     return d;
191 }
192 
193 template <typename D, typename S>
194 SI D bit_pun(const S& s) {
195     static_assert(sizeof(D) == sizeof(S), "");
196     return unchecked_bit_pun<D>(s);
197 }
198 
199 // Translate from a value type T to its corresponding Mask, the result of a comparison.
200 template <typename T> struct Mask { using type = T; };
201 template <> struct Mask<float > { using type = int32_t; };
202 template <> struct Mask<double> { using type = int64_t; };
203 template <typename T> using M = typename Mask<T>::type;
204 
205 // Join two Vec<N,T> into one Vec<2N,T>.
206 SINT Vec<2*N,T> join(const Vec<N,T>& lo, const Vec<N,T>& hi) {
207     Vec<2*N,T> v;
208     v.lo = lo;
209     v.hi = hi;
210     return v;
211 }
212 
213 // We have three strategies for implementing Vec operations:
214 //    1) lean on Clang/GCC vector extensions when available;
215 //    2) use map() to apply a scalar function lane-wise;
216 //    3) recurse on lo/hi to scalar portable implementations.
217 // We can slot in platform-specific implementations as overloads for particular Vec<N,T>,
218 // or often integrate them directly into the recursion of style 3), allowing fine control.
219 
220 #if !defined(SKNX_NO_SIMD) && (defined(__clang__) || defined(__GNUC__))
221 
222     // VExt<N,T> types have the same size as Vec<N,T> and support most operations directly.
223     #if defined(__clang__)
224         template <int N, typename T>
225         using VExt = T __attribute__((ext_vector_type(N)));
226 
227     #elif defined(__GNUC__)
228         template <int N, typename T>
229         struct VExtHelper {
230             typedef T __attribute__((vector_size(N*sizeof(T)))) type;
231         };
232 
233         template <int N, typename T>
234         using VExt = typename VExtHelper<N,T>::type;
235 
236         // For some reason some (new!) versions of GCC cannot seem to deduce N in the generic
237         // to_vec<N,T>() below for N=4 and T=float.  This workaround seems to help...
238         SI Vec<4,float> to_vec(VExt<4,float> v) { return bit_pun<Vec<4,float>>(v); }
239     #endif
240 
241     SINT VExt<N,T> to_vext(const Vec<N,T>& v) { return bit_pun<VExt<N,T>>(v); }
242     SINT Vec <N,T> to_vec(const VExt<N,T>& v) { return bit_pun<Vec <N,T>>(v); }
243 
244     SINT Vec<N,T> operator+(const Vec<N,T>& x, const Vec<N,T>& y) {
245         return to_vec<N,T>(to_vext(x) + to_vext(y));
246     }
247     SINT Vec<N,T> operator-(const Vec<N,T>& x, const Vec<N,T>& y) {
248         return to_vec<N,T>(to_vext(x) - to_vext(y));
249     }
250     SINT Vec<N,T> operator*(const Vec<N,T>& x, const Vec<N,T>& y) {
251         return to_vec<N,T>(to_vext(x) * to_vext(y));
252     }
253     SINT Vec<N,T> operator/(const Vec<N,T>& x, const Vec<N,T>& y) {
254         return to_vec<N,T>(to_vext(x) / to_vext(y));
255     }
256 
257     SINT Vec<N,T> operator^(const Vec<N,T>& x, const Vec<N,T>& y) {
258         return to_vec<N,T>(to_vext(x) ^ to_vext(y));
259     }
260     SINT Vec<N,T> operator&(const Vec<N,T>& x, const Vec<N,T>& y) {
261         return to_vec<N,T>(to_vext(x) & to_vext(y));
262     }
263     SINT Vec<N,T> operator|(const Vec<N,T>& x, const Vec<N,T>& y) {
264         return to_vec<N,T>(to_vext(x) | to_vext(y));
265     }
266 
267     SINT Vec<N,T> operator!(const Vec<N,T>& x) { return to_vec<N,T>(!to_vext(x)); }
268     SINT Vec<N,T> operator-(const Vec<N,T>& x) { return to_vec<N,T>(-to_vext(x)); }
269     SINT Vec<N,T> operator~(const Vec<N,T>& x) { return to_vec<N,T>(~to_vext(x)); }
270 
271     SINT Vec<N,T> operator<<(const Vec<N,T>& x, int k) { return to_vec<N,T>(to_vext(x) << k); }
272     SINT Vec<N,T> operator>>(const Vec<N,T>& x, int k) { return to_vec<N,T>(to_vext(x) >> k); }
273 
274     SINT Vec<N,M<T>> operator==(const Vec<N,T>& x, const Vec<N,T>& y) {
275         return bit_pun<Vec<N,M<T>>>(to_vext(x) == to_vext(y));
276     }
277     SINT Vec<N,M<T>> operator!=(const Vec<N,T>& x, const Vec<N,T>& y) {
278         return bit_pun<Vec<N,M<T>>>(to_vext(x) != to_vext(y));
279     }
280     SINT Vec<N,M<T>> operator<=(const Vec<N,T>& x, const Vec<N,T>& y) {
281         return bit_pun<Vec<N,M<T>>>(to_vext(x) <= to_vext(y));
282     }
283     SINT Vec<N,M<T>> operator>=(const Vec<N,T>& x, const Vec<N,T>& y) {
284         return bit_pun<Vec<N,M<T>>>(to_vext(x) >= to_vext(y));
285     }
286     SINT Vec<N,M<T>> operator< (const Vec<N,T>& x, const Vec<N,T>& y) {
287         return bit_pun<Vec<N,M<T>>>(to_vext(x) <  to_vext(y));
288     }
289     SINT Vec<N,M<T>> operator> (const Vec<N,T>& x, const Vec<N,T>& y) {
290         return bit_pun<Vec<N,M<T>>>(to_vext(x) >  to_vext(y));
291     }
292 
293 #else
294 
295     // Either SKNX_NO_SIMD is defined, or Clang/GCC vector extensions are not available.
296     // We'll implement things portably with N==1 scalar implementations and recursion onto them.
297 
298     // N == 1 scalar implementations.
299     SIT Vec<1,T> operator+(const Vec<1,T>& x, const Vec<1,T>& y) { return x.val + y.val; }
300     SIT Vec<1,T> operator-(const Vec<1,T>& x, const Vec<1,T>& y) { return x.val - y.val; }
301     SIT Vec<1,T> operator*(const Vec<1,T>& x, const Vec<1,T>& y) { return x.val * y.val; }
302     SIT Vec<1,T> operator/(const Vec<1,T>& x, const Vec<1,T>& y) { return x.val / y.val; }
303 
304     SIT Vec<1,T> operator^(const Vec<1,T>& x, const Vec<1,T>& y) { return x.val ^ y.val; }
305     SIT Vec<1,T> operator&(const Vec<1,T>& x, const Vec<1,T>& y) { return x.val & y.val; }
306     SIT Vec<1,T> operator|(const Vec<1,T>& x, const Vec<1,T>& y) { return x.val | y.val; }
307 
308     SIT Vec<1,T> operator!(const Vec<1,T>& x) { return !x.val; }
309     SIT Vec<1,T> operator-(const Vec<1,T>& x) { return -x.val; }
310     SIT Vec<1,T> operator~(const Vec<1,T>& x) { return ~x.val; }
311 
312     SIT Vec<1,T> operator<<(const Vec<1,T>& x, int k) { return x.val << k; }
313     SIT Vec<1,T> operator>>(const Vec<1,T>& x, int k) { return x.val >> k; }
314 
315     SIT Vec<1,M<T>> operator==(const Vec<1,T>& x, const Vec<1,T>& y) {
316         return x.val == y.val ? ~0 : 0;
317     }
318     SIT Vec<1,M<T>> operator!=(const Vec<1,T>& x, const Vec<1,T>& y) {
319         return x.val != y.val ? ~0 : 0;
320     }
321     SIT Vec<1,M<T>> operator<=(const Vec<1,T>& x, const Vec<1,T>& y) {
322         return x.val <= y.val ? ~0 : 0;
323     }
324     SIT Vec<1,M<T>> operator>=(const Vec<1,T>& x, const Vec<1,T>& y) {
325         return x.val >= y.val ? ~0 : 0;
326     }
327     SIT Vec<1,M<T>> operator< (const Vec<1,T>& x, const Vec<1,T>& y) {
328         return x.val <  y.val ? ~0 : 0;
329     }
330     SIT Vec<1,M<T>> operator> (const Vec<1,T>& x, const Vec<1,T>& y) {
331         return x.val >  y.val ? ~0 : 0;
332     }
333 
334     // Recurse on lo/hi down to N==1 scalar implementations.
335     SINT Vec<N,T> operator+(const Vec<N,T>& x, const Vec<N,T>& y) {
336         return join(x.lo + y.lo, x.hi + y.hi);
337     }
338     SINT Vec<N,T> operator-(const Vec<N,T>& x, const Vec<N,T>& y) {
339         return join(x.lo - y.lo, x.hi - y.hi);
340     }
341     SINT Vec<N,T> operator*(const Vec<N,T>& x, const Vec<N,T>& y) {
342         return join(x.lo * y.lo, x.hi * y.hi);
343     }
344     SINT Vec<N,T> operator/(const Vec<N,T>& x, const Vec<N,T>& y) {
345         return join(x.lo / y.lo, x.hi / y.hi);
346     }
347 
348     SINT Vec<N,T> operator^(const Vec<N,T>& x, const Vec<N,T>& y) {
349         return join(x.lo ^ y.lo, x.hi ^ y.hi);
350     }
351     SINT Vec<N,T> operator&(const Vec<N,T>& x, const Vec<N,T>& y) {
352         return join(x.lo & y.lo, x.hi & y.hi);
353     }
354     SINT Vec<N,T> operator|(const Vec<N,T>& x, const Vec<N,T>& y) {
355         return join(x.lo | y.lo, x.hi | y.hi);
356     }
357 
358     SINT Vec<N,T> operator!(const Vec<N,T>& x) { return join(!x.lo, !x.hi); }
359     SINT Vec<N,T> operator-(const Vec<N,T>& x) { return join(-x.lo, -x.hi); }
360     SINT Vec<N,T> operator~(const Vec<N,T>& x) { return join(~x.lo, ~x.hi); }
361 
362     SINT Vec<N,T> operator<<(const Vec<N,T>& x, int k) { return join(x.lo << k, x.hi << k); }
363     SINT Vec<N,T> operator>>(const Vec<N,T>& x, int k) { return join(x.lo >> k, x.hi >> k); }
364 
365     SINT Vec<N,M<T>> operator==(const Vec<N,T>& x, const Vec<N,T>& y) {
366         return join(x.lo == y.lo, x.hi == y.hi);
367     }
368     SINT Vec<N,M<T>> operator!=(const Vec<N,T>& x, const Vec<N,T>& y) {
369         return join(x.lo != y.lo, x.hi != y.hi);
370     }
371     SINT Vec<N,M<T>> operator<=(const Vec<N,T>& x, const Vec<N,T>& y) {
372         return join(x.lo <= y.lo, x.hi <= y.hi);
373     }
374     SINT Vec<N,M<T>> operator>=(const Vec<N,T>& x, const Vec<N,T>& y) {
375         return join(x.lo >= y.lo, x.hi >= y.hi);
376     }
377     SINT Vec<N,M<T>> operator< (const Vec<N,T>& x, const Vec<N,T>& y) {
378         return join(x.lo <  y.lo, x.hi <  y.hi);
379     }
380     SINT Vec<N,M<T>> operator> (const Vec<N,T>& x, const Vec<N,T>& y) {
381         return join(x.lo >  y.lo, x.hi >  y.hi);
382     }
383 #endif
384 
385 // Scalar/vector operations splat the scalar to a vector.
386 SINTU Vec<N,T>    operator+ (U x, const Vec<N,T>& y) { return Vec<N,T>(x) +  y; }
387 SINTU Vec<N,T>    operator- (U x, const Vec<N,T>& y) { return Vec<N,T>(x) -  y; }
388 SINTU Vec<N,T>    operator* (U x, const Vec<N,T>& y) { return Vec<N,T>(x) *  y; }
389 SINTU Vec<N,T>    operator/ (U x, const Vec<N,T>& y) { return Vec<N,T>(x) /  y; }
390 SINTU Vec<N,T>    operator^ (U x, const Vec<N,T>& y) { return Vec<N,T>(x) ^  y; }
391 SINTU Vec<N,T>    operator& (U x, const Vec<N,T>& y) { return Vec<N,T>(x) &  y; }
392 SINTU Vec<N,T>    operator| (U x, const Vec<N,T>& y) { return Vec<N,T>(x) |  y; }
393 SINTU Vec<N,M<T>> operator==(U x, const Vec<N,T>& y) { return Vec<N,T>(x) == y; }
394 SINTU Vec<N,M<T>> operator!=(U x, const Vec<N,T>& y) { return Vec<N,T>(x) != y; }
395 SINTU Vec<N,M<T>> operator<=(U x, const Vec<N,T>& y) { return Vec<N,T>(x) <= y; }
396 SINTU Vec<N,M<T>> operator>=(U x, const Vec<N,T>& y) { return Vec<N,T>(x) >= y; }
397 SINTU Vec<N,M<T>> operator< (U x, const Vec<N,T>& y) { return Vec<N,T>(x) <  y; }
398 SINTU Vec<N,M<T>> operator> (U x, const Vec<N,T>& y) { return Vec<N,T>(x) >  y; }
399 
400 SINTU Vec<N,T>    operator+ (const Vec<N,T>& x, U y) { return x +  Vec<N,T>(y); }
401 SINTU Vec<N,T>    operator- (const Vec<N,T>& x, U y) { return x -  Vec<N,T>(y); }
402 SINTU Vec<N,T>    operator* (const Vec<N,T>& x, U y) { return x *  Vec<N,T>(y); }
403 SINTU Vec<N,T>    operator/ (const Vec<N,T>& x, U y) { return x /  Vec<N,T>(y); }
404 SINTU Vec<N,T>    operator^ (const Vec<N,T>& x, U y) { return x ^  Vec<N,T>(y); }
405 SINTU Vec<N,T>    operator& (const Vec<N,T>& x, U y) { return x &  Vec<N,T>(y); }
406 SINTU Vec<N,T>    operator| (const Vec<N,T>& x, U y) { return x |  Vec<N,T>(y); }
407 SINTU Vec<N,M<T>> operator==(const Vec<N,T>& x, U y) { return x == Vec<N,T>(y); }
408 SINTU Vec<N,M<T>> operator!=(const Vec<N,T>& x, U y) { return x != Vec<N,T>(y); }
409 SINTU Vec<N,M<T>> operator<=(const Vec<N,T>& x, U y) { return x <= Vec<N,T>(y); }
410 SINTU Vec<N,M<T>> operator>=(const Vec<N,T>& x, U y) { return x >= Vec<N,T>(y); }
411 SINTU Vec<N,M<T>> operator< (const Vec<N,T>& x, U y) { return x <  Vec<N,T>(y); }
412 SINTU Vec<N,M<T>> operator> (const Vec<N,T>& x, U y) { return x >  Vec<N,T>(y); }
413 
414 SINT Vec<N,T>& operator+=(Vec<N,T>& x, const Vec<N,T>& y) { return (x = x + y); }
415 SINT Vec<N,T>& operator-=(Vec<N,T>& x, const Vec<N,T>& y) { return (x = x - y); }
416 SINT Vec<N,T>& operator*=(Vec<N,T>& x, const Vec<N,T>& y) { return (x = x * y); }
417 SINT Vec<N,T>& operator/=(Vec<N,T>& x, const Vec<N,T>& y) { return (x = x / y); }
418 SINT Vec<N,T>& operator^=(Vec<N,T>& x, const Vec<N,T>& y) { return (x = x ^ y); }
419 SINT Vec<N,T>& operator&=(Vec<N,T>& x, const Vec<N,T>& y) { return (x = x & y); }
420 SINT Vec<N,T>& operator|=(Vec<N,T>& x, const Vec<N,T>& y) { return (x = x | y); }
421 
422 SINTU Vec<N,T>& operator+=(Vec<N,T>& x, U y) { return (x = x + Vec<N,T>(y)); }
423 SINTU Vec<N,T>& operator-=(Vec<N,T>& x, U y) { return (x = x - Vec<N,T>(y)); }
424 SINTU Vec<N,T>& operator*=(Vec<N,T>& x, U y) { return (x = x * Vec<N,T>(y)); }
425 SINTU Vec<N,T>& operator/=(Vec<N,T>& x, U y) { return (x = x / Vec<N,T>(y)); }
426 SINTU Vec<N,T>& operator^=(Vec<N,T>& x, U y) { return (x = x ^ Vec<N,T>(y)); }
427 SINTU Vec<N,T>& operator&=(Vec<N,T>& x, U y) { return (x = x & Vec<N,T>(y)); }
428 SINTU Vec<N,T>& operator|=(Vec<N,T>& x, U y) { return (x = x | Vec<N,T>(y)); }
429 
430 SINT Vec<N,T>& operator<<=(Vec<N,T>& x, int bits) { return (x = x << bits); }
431 SINT Vec<N,T>& operator>>=(Vec<N,T>& x, int bits) { return (x = x >> bits); }
432 
433 // Some operations we want are not expressible with Clang/GCC vector extensions.
434 
435 // Clang can reason about naive_if_then_else() and optimize through it better
436 // than if_then_else(), so it's sometimes useful to call it directly when we
437 // think an entire expression should optimize away, e.g. min()/max().
438 SINT Vec<N,T> naive_if_then_else(const Vec<N,M<T>>& cond, const Vec<N,T>& t, const Vec<N,T>& e) {
439     return bit_pun<Vec<N,T>>(( cond & bit_pun<Vec<N, M<T>>>(t)) |
440                              (~cond & bit_pun<Vec<N, M<T>>>(e)) );
441 }
442 
443 SIT Vec<1,T> if_then_else(const Vec<1,M<T>>& cond, const Vec<1,T>& t, const Vec<1,T>& e) {
444     // In practice this scalar implementation is unlikely to be used.  See next if_then_else().
445     return bit_pun<Vec<1,T>>(( cond & bit_pun<Vec<1, M<T>>>(t)) |
446                              (~cond & bit_pun<Vec<1, M<T>>>(e)) );
447 }
448 SINT Vec<N,T> if_then_else(const Vec<N,M<T>>& cond, const Vec<N,T>& t, const Vec<N,T>& e) {
449     // Specializations inline here so they can generalize what types the apply to.
450     // (This header is used in C++14 contexts, so we have to kind of fake constexpr if.)
451 #if defined(__AVX2__)
452     if /*constexpr*/ (N*sizeof(T) == 32) {
453         return unchecked_bit_pun<Vec<N,T>>(_mm256_blendv_epi8(unchecked_bit_pun<__m256i>(e),
454                                                               unchecked_bit_pun<__m256i>(t),
455                                                               unchecked_bit_pun<__m256i>(cond)));
456     }
457 #endif
458 #if defined(__SSE4_1__)
459     if /*constexpr*/ (N*sizeof(T) == 16) {
460         return unchecked_bit_pun<Vec<N,T>>(_mm_blendv_epi8(unchecked_bit_pun<__m128i>(e),
461                                                            unchecked_bit_pun<__m128i>(t),
462                                                            unchecked_bit_pun<__m128i>(cond)));
463     }
464 #endif
465 #if defined(__ARM_NEON)
466     if /*constexpr*/ (N*sizeof(T) == 16) {
467         return unchecked_bit_pun<Vec<N,T>>(vbslq_u8(unchecked_bit_pun<uint8x16_t>(cond),
468                                                     unchecked_bit_pun<uint8x16_t>(t),
469                                                     unchecked_bit_pun<uint8x16_t>(e)));
470     }
471 #endif
472     // Recurse for large vectors to try to hit the specializations above.
473     if /*constexpr*/ (N*sizeof(T) > 16) {
474         return join(if_then_else(cond.lo, t.lo, e.lo),
475                     if_then_else(cond.hi, t.hi, e.hi));
476     }
477     // This default can lead to better code than the recursing onto scalars.
478     return naive_if_then_else(cond, t, e);
479 }
480 
481 SIT  bool any(const Vec<1,T>& x) { return x.val != 0; }
482 SINT bool any(const Vec<N,T>& x) {
483 #if defined(__wasm_simd128__)
484     if constexpr (N == 4 && sizeof(T) == 4) {
485         return wasm_i32x4_any_true(unchecked_bit_pun<VExt<4,int>>(x));
486     }
487 #endif
488     return any(x.lo)
489         || any(x.hi);
490 }
491 
492 SIT  bool all(const Vec<1,T>& x) { return x.val != 0; }
493 SINT bool all(const Vec<N,T>& x) {
494 #if defined(__AVX2__)
495     if /*constexpr*/ (N*sizeof(T) == 32) {
496         return _mm256_testc_si256(unchecked_bit_pun<__m256i>(x),
497                                   _mm256_set1_epi32(-1));
498     }
499 #endif
500 #if defined(__SSE4_1__)
501     if /*constexpr*/ (N*sizeof(T) == 16) {
502         return _mm_testc_si128(unchecked_bit_pun<__m128i>(x),
503                                _mm_set1_epi32(-1));
504     }
505 #endif
506 #if defined(__wasm_simd128__)
507     if /*constexpr*/ (N == 4 && sizeof(T) == 4) {
508         return wasm_i32x4_all_true(unchecked_bit_pun<VExt<4,int>>(x));
509     }
510 #endif
511     return all(x.lo)
512         && all(x.hi);
513 }
514 
515 // cast() Vec<N,S> to Vec<N,D>, as if applying a C-cast to each lane.
516 // TODO: implement with map()?
517 template <typename D, typename S>
518 SI Vec<1,D> cast(const Vec<1,S>& src) { return (D)src.val; }
519 
520 template <typename D, int N, typename S>
521 SI Vec<N,D> cast(const Vec<N,S>& src) {
522 #if !defined(SKNX_NO_SIMD) && defined(__clang__)
523     return to_vec(__builtin_convertvector(to_vext(src), VExt<N,D>));
524 #else
525     return join(cast<D>(src.lo), cast<D>(src.hi));
526 #endif
527 }
528 
529 // min/max match logic of std::min/std::max, which is important when NaN is involved.
530 SIT  T min(const Vec<1,T>& x) { return x.val; }
531 SIT  T max(const Vec<1,T>& x) { return x.val; }
532 SINT T min(const Vec<N,T>& x) { return std::min(min(x.lo), min(x.hi)); }
533 SINT T max(const Vec<N,T>& x) { return std::max(max(x.lo), max(x.hi)); }
534 
535 SINT Vec<N,T> min(const Vec<N,T>& x, const Vec<N,T>& y) { return naive_if_then_else(y < x, y, x); }
536 SINT Vec<N,T> max(const Vec<N,T>& x, const Vec<N,T>& y) { return naive_if_then_else(x < y, y, x); }
537 
538 SINTU Vec<N,T> min(const Vec<N,T>& x, U y) { return min(x, Vec<N,T>(y)); }
539 SINTU Vec<N,T> max(const Vec<N,T>& x, U y) { return max(x, Vec<N,T>(y)); }
540 SINTU Vec<N,T> min(U x, const Vec<N,T>& y) { return min(Vec<N,T>(x), y); }
541 SINTU Vec<N,T> max(U x, const Vec<N,T>& y) { return max(Vec<N,T>(x), y); }
542 
543 // pin matches the logic of SkTPin, which is important when NaN is involved. It always returns
544 // values in the range lo..hi, and if x is NaN, it returns lo.
545 SINT Vec<N,T> pin(const Vec<N,T>& x, const Vec<N,T>& lo, const Vec<N,T>& hi) {
546     return max(lo, min(x, hi));
547 }
548 
549 // Shuffle values from a vector pretty arbitrarily:
550 //    skvx::Vec<4,float> rgba = {R,G,B,A};
551 //    shuffle<2,1,0,3>        (rgba) ~> {B,G,R,A}
552 //    shuffle<2,1>            (rgba) ~> {B,G}
553 //    shuffle<2,1,2,1,2,1,2,1>(rgba) ~> {B,G,B,G,B,G,B,G}
554 //    shuffle<3,3,3,3>        (rgba) ~> {A,A,A,A}
555 // The only real restriction is that the output also be a legal N=power-of-two sknx::Vec.
556 template <int... Ix, int N, typename T>
557 SI Vec<sizeof...(Ix),T> shuffle(const Vec<N,T>& x) {
558 #if !defined(SKNX_NO_SIMD) && defined(__clang__)
559     // TODO: can we just always use { x[Ix]... }?
560     return to_vec<sizeof...(Ix),T>(__builtin_shufflevector(to_vext(x), to_vext(x), Ix...));
561 #else
562     return { x[Ix]... };
563 #endif
564 }
565 
566 // Call map(fn, x) for a vector with fn() applied to each lane of x, { fn(x[0]), fn(x[1]), ... },
567 // or map(fn, x,y) for a vector of fn(x[i], y[i]), etc.
568 
569 template <typename Fn, typename... Args, size_t... I>
570 SI auto map(std::index_sequence<I...>,
571             Fn&& fn, const Args&... args) -> skvx::Vec<sizeof...(I), decltype(fn(args[0]...))> {
572     auto lane = [&](size_t i)
573 #if defined(__clang__)
574     // CFI, specifically -fsanitize=cfi-icall, seems to give a false positive here,
575     // with errors like "control flow integrity check for type 'float (float)
576     // noexcept' failed during indirect function call... note: sqrtf.cfi_jt defined
577     // here".  But we can be quite sure fn is the right type: it's all inferred!
578     // So, stifle CFI in this function.
579     __attribute__((no_sanitize("cfi")))
580 #endif
581     { return fn(args[i]...); };
582 
583     return { lane(I)... };
584 }
585 
586 template <typename Fn, int N, typename T, typename... Rest>
587 auto map(Fn&& fn, const Vec<N,T>& first, const Rest&... rest) {
588     // Derive an {0...N-1} index_sequence from the size of the first arg: N lanes in, N lanes out.
589     return map(std::make_index_sequence<N>{}, fn, first,rest...);
590 }
591 
592 SIN Vec<N,float>  ceil(const Vec<N,float>& x) { return map( ceilf, x); }
593 SIN Vec<N,float> floor(const Vec<N,float>& x) { return map(floorf, x); }
594 SIN Vec<N,float> trunc(const Vec<N,float>& x) { return map(truncf, x); }
595 SIN Vec<N,float> round(const Vec<N,float>& x) { return map(roundf, x); }
596 SIN Vec<N,float>  sqrt(const Vec<N,float>& x) { return map( sqrtf, x); }
597 SIN Vec<N,float>   abs(const Vec<N,float>& x) { return map( fabsf, x); }
598 SIN Vec<N,float>   fma(const Vec<N,float>& x,
599                        const Vec<N,float>& y,
600                        const Vec<N,float>& z) {
601     // I don't understand why Clang's codegen is terrible if we write map(fmaf, x,y,z) directly.
602     auto fn = [](float x, float y, float z) { return fmaf(x,y,z); };
603     return map(fn, x,y,z);
604 }
605 
606 SI Vec<1,int> lrint(const Vec<1,float>& x) {
607     return (int)lrintf(x.val);
608 }
609 SIN Vec<N,int> lrint(const Vec<N,float>& x) {
610 #if defined(__AVX__)
611     if /*constexpr*/ (N == 8) {
612         return unchecked_bit_pun<Vec<N,int>>(_mm256_cvtps_epi32(unchecked_bit_pun<__m256>(x)));
613     }
614 #endif
615 #if defined(__SSE__)
616     if /*constexpr*/ (N == 4) {
617         return unchecked_bit_pun<Vec<N,int>>(_mm_cvtps_epi32(unchecked_bit_pun<__m128>(x)));
618     }
619 #endif
620     return join(lrint(x.lo),
621                 lrint(x.hi));
622 }
623 
624 SIN Vec<N,float> fract(const Vec<N,float>& x) { return x - floor(x); }
625 
626 // The default logic for to_half/from_half is borrowed from skcms,
627 // and assumes inputs are finite and treat/flush denorm half floats as/to zero.
628 // Key constants to watch for:
629 //    - a float is 32-bit, 1-8-23 sign-exponent-mantissa, with 127 exponent bias;
630 //    - a half  is 16-bit, 1-5-10 sign-exponent-mantissa, with  15 exponent bias.
631 SIN Vec<N,uint16_t> to_half_finite_ftz(const Vec<N,float>& x) {
632     Vec<N,uint32_t> sem = bit_pun<Vec<N,uint32_t>>(x),
633                     s   = sem & 0x8000'0000,
634                      em = sem ^ s,
635               is_denorm =  em < 0x3880'0000;
636     return cast<uint16_t>(if_then_else(is_denorm, Vec<N,uint32_t>(0)
637                                                 , (s>>16) + (em>>13) - ((127-15)<<10)));
638 }
639 SIN Vec<N,float> from_half_finite_ftz(const Vec<N,uint16_t>& x) {
640     Vec<N,uint32_t> wide = cast<uint32_t>(x),
641                       s  = wide & 0x8000,
642                       em = wide ^ s;
643     auto is_denorm = bit_pun<Vec<N,int32_t>>(em < 0x0400);
644     return if_then_else(is_denorm, Vec<N,float>(0)
645                                  , bit_pun<Vec<N,float>>( (s<<16) + (em<<13) + ((127-15)<<23) ));
646 }
647 
648 // Like if_then_else(), these N=1 base cases won't actually be used unless explicitly called.
649 SI Vec<1,uint16_t> to_half(const Vec<1,float>&    x) { return   to_half_finite_ftz(x); }
650 SI Vec<1,float>  from_half(const Vec<1,uint16_t>& x) { return from_half_finite_ftz(x); }
651 
652 SIN Vec<N,uint16_t> to_half(const Vec<N,float>& x) {
653 #if defined(__F16C__)
654     if /*constexpr*/ (N == 8) {
655         return unchecked_bit_pun<Vec<N,uint16_t>>(_mm256_cvtps_ph(unchecked_bit_pun<__m256>(x),
656                                                                   _MM_FROUND_CUR_DIRECTION));
657     }
658 #endif
659 #if defined(__aarch64__)
660     if /*constexpr*/ (N == 4) {
661         return unchecked_bit_pun<Vec<N,uint16_t>>(vcvt_f16_f32(unchecked_bit_pun<float32x4_t>(x)));
662 
663     }
664 #endif
665     if /*constexpr*/ (N > 4) {
666         return join(to_half(x.lo),
667                     to_half(x.hi));
668     }
669     return to_half_finite_ftz(x);
670 }
671 
672 SIN Vec<N,float> from_half(const Vec<N,uint16_t>& x) {
673 #if defined(__F16C__)
674     if /*constexpr*/ (N == 8) {
675         return unchecked_bit_pun<Vec<N,float>>(_mm256_cvtph_ps(unchecked_bit_pun<__m128i>(x)));
676     }
677 #endif
678 #if defined(__aarch64__)
679     if /*constexpr*/ (N == 4) {
680         return unchecked_bit_pun<Vec<N,float>>(vcvt_f32_f16(unchecked_bit_pun<float16x4_t>(x)));
681     }
682 #endif
683     if /*constexpr*/ (N > 4) {
684         return join(from_half(x.lo),
685                     from_half(x.hi));
686     }
687     return from_half_finite_ftz(x);
688 }
689 
690 // div255(x) = (x + 127) / 255 is a bit-exact rounding divide-by-255, packing down to 8-bit.
691 SIN Vec<N,uint8_t> div255(const Vec<N,uint16_t>& x) {
692     return cast<uint8_t>( (x+127)/255 );
693 }
694 
695 // approx_scale(x,y) approximates div255(cast<uint16_t>(x)*cast<uint16_t>(y)) within a bit,
696 // and is always perfect when x or y is 0 or 255.
697 SIN Vec<N,uint8_t> approx_scale(const Vec<N,uint8_t>& x, const Vec<N,uint8_t>& y) {
698     // All of (x*y+x)/256, (x*y+y)/256, and (x*y+255)/256 meet the criteria above.
699     // We happen to have historically picked (x*y+x)/256.
700     auto X = cast<uint16_t>(x),
701          Y = cast<uint16_t>(y);
702     return cast<uint8_t>( (X*Y+X)/256 );
703 }
704 
705 // The ScaledDividerU32 takes a divisor > 1, and creates a function divide(numerator) that
706 // calculates a numerator / denominator. For this to be rounded properly, numerator should have
707 // half added in:
708 // divide(numerator + half) == floor(numerator/denominator + 1/2).
709 //
710 // This gives an answer within +/- 1 from the true value.
711 //
712 // Derivation of half:
713 //    numerator/denominator + 1/2 = (numerator + half) / d
714 //    numerator + denominator / 2 = numerator + half
715 //    half = denominator / 2.
716 //
717 // Because half is divided by 2, that division must also be rounded.
718 //    half == denominator / 2 = (denominator + 1) / 2.
719 //
720 // The divisorFactor is just a scaled value:
721 //    divisorFactor = (1 / divisor) * 2 ^ 32.
722 // The maximum that can be divided and rounded is UINT_MAX - half.
723 class ScaledDividerU32 {
724 public:
725     explicit ScaledDividerU32(uint32_t divisor)
726             : fDivisorFactor{(uint32_t)(std::round((1.0 / divisor) * (1ull << 32)))}
727             , fHalf{(divisor + 1) >> 1} {
728         assert(divisor > 1);
729     }
730 
731     Vec<4, uint32_t> divide(const Vec<4, uint32_t>& numerator) const {
732     #if !defined(SKNX_NO_SIMD) && defined(__ARM_NEON)
733         uint64x2_t hi = vmull_n_u32(vget_high_u32(to_vext(numerator)), fDivisorFactor);
734         uint64x2_t lo = vmull_n_u32(vget_low_u32(to_vext(numerator)),  fDivisorFactor);
735 
736         return to_vec<4, uint32_t>(vcombine_u32(vshrn_n_u64(lo,32), vshrn_n_u64(hi,32)));
737     #else
738         return cast<uint32_t>((cast<uint64_t>(numerator) * fDivisorFactor) >> 32);
739     #endif
740     }
741 
742     uint32_t half() const { return fHalf; }
743 
744 private:
745     const uint32_t fDivisorFactor;
746     const uint32_t fHalf;
747 };
748 
749 #if !defined(SKNX_NO_SIMD) && defined(__ARM_NEON)
750     // With NEON we can do eight u8*u8 -> u16 in one instruction, vmull_u8 (read, mul-long).
751     SI Vec<8,uint16_t> mull(const Vec<8,uint8_t>& x,
752                             const Vec<8,uint8_t>& y) {
753         return to_vec<8,uint16_t>(vmull_u8(to_vext(x),
754                                            to_vext(y)));
755     }
756 
757     SIN std::enable_if_t<(N < 8), Vec<N,uint16_t>> mull(const Vec<N,uint8_t>& x,
758                                                         const Vec<N,uint8_t>& y) {
759         // N < 8 --> double up data until N == 8, returning the part we need.
760         return mull(join(x,x),
761                     join(y,y)).lo;
762     }
763 
764     SIN std::enable_if_t<(N > 8), Vec<N,uint16_t>> mull(const Vec<N,uint8_t>& x,
765                                                         const Vec<N,uint8_t>& y) {
766         // N > 8 --> usual join(lo,hi) strategy to recurse down to N == 8.
767         return join(mull(x.lo, y.lo),
768                     mull(x.hi, y.hi));
769     }
770 #else
771     // Nothing special when we don't have NEON... just cast up to 16-bit and multiply.
772     SIN Vec<N,uint16_t> mull(const Vec<N,uint8_t>& x,
773                              const Vec<N,uint8_t>& y) {
774         return cast<uint16_t>(x)
775              * cast<uint16_t>(y);
776     }
777 #endif
778 
779 // Allow floating point contraction. e.g., allow a*x + y to be compiled to a single FMA even though
780 // it introduces LSB differences on platforms that don't have an FMA instruction.
781 #if defined(__clang__)
782 #pragma STDC FP_CONTRACT ON
783 #endif
784 
785 // Approximates the inverse cosine of x within 0.96 degrees using the rational polynomial:
786 //
787 //     acos(x) ~= (bx^3 + ax) / (dx^4 + cx^2 + 1) + pi/2
788 //
789 // See: https://stackoverflow.com/a/36387954
790 //
791 // For a proof of max error, see the "SkVx_approx_acos" unit test.
792 //
793 // NOTE: This function deviates immediately from pi and 0 outside -1 and 1. (The derivatives are
794 // infinite at -1 and 1). So the input must still be clamped between -1 and 1.
795 #define SKVX_APPROX_ACOS_MAX_ERROR SkDegreesToRadians(.96f)
796 SIN Vec<N,float> approx_acos(Vec<N,float> x) {
797     constexpr static float a = -0.939115566365855f;
798     constexpr static float b =  0.9217841528914573f;
799     constexpr static float c = -1.2845906244690837f;
800     constexpr static float d =  0.295624144969963174f;
801     constexpr static float pi_over_2 = 1.5707963267948966f;
802     auto xx = x*x;
803     auto numer = b*xx + a;
804     auto denom = xx*(d*xx + c) + 1;
805     return x * (numer/denom) + pi_over_2;
806 }
807 
808 #if defined(__clang__)
809 #pragma STDC FP_CONTRACT DEFAULT
810 #endif
811 
812 // De-interleaving load of 4 vectors.
813 //
814 // WARNING: These are really only supported well on NEON. Consider restructuring your data before
815 // resorting to these methods.
816 SIT void strided_load4(const T* v,
817                        skvx::Vec<1,T>& a,
818                        skvx::Vec<1,T>& b,
819                        skvx::Vec<1,T>& c,
820                        skvx::Vec<1,T>& d) {
821     a.val = v[0];
822     b.val = v[1];
823     c.val = v[2];
824     d.val = v[3];
825 }
826 SINT void strided_load4(const T* v,
827                         skvx::Vec<N,T>& a,
828                         skvx::Vec<N,T>& b,
829                         skvx::Vec<N,T>& c,
830                         skvx::Vec<N,T>& d) {
831     strided_load4(v, a.lo, b.lo, c.lo, d.lo);
832     strided_load4(v + 4*(N/2), a.hi, b.hi, c.hi, d.hi);
833 }
834 #if !defined(SKNX_NO_SIMD)
835 #if defined(__ARM_NEON)
836 #define IMPL_LOAD4_TRANSPOSED(N, T, VLD) \
837 SI void strided_load4(const T* v, \
838                       skvx::Vec<N,T>& a, \
839                       skvx::Vec<N,T>& b, \
840                       skvx::Vec<N,T>& c, \
841                       skvx::Vec<N,T>& d) { \
842     auto mat = VLD(v); \
843     a = skvx::bit_pun<skvx::Vec<N,T>>(mat.val[0]); \
844     b = skvx::bit_pun<skvx::Vec<N,T>>(mat.val[1]); \
845     c = skvx::bit_pun<skvx::Vec<N,T>>(mat.val[2]); \
846     d = skvx::bit_pun<skvx::Vec<N,T>>(mat.val[3]); \
847 }
848 IMPL_LOAD4_TRANSPOSED(2, uint32_t, vld4_u32);
849 IMPL_LOAD4_TRANSPOSED(4, uint16_t, vld4_u16);
850 IMPL_LOAD4_TRANSPOSED(8, uint8_t, vld4_u8);
851 IMPL_LOAD4_TRANSPOSED(2, int32_t, vld4_s32);
852 IMPL_LOAD4_TRANSPOSED(4, int16_t, vld4_s16);
853 IMPL_LOAD4_TRANSPOSED(8, int8_t, vld4_s8);
854 IMPL_LOAD4_TRANSPOSED(2, float, vld4_f32);
855 IMPL_LOAD4_TRANSPOSED(4, uint32_t, vld4q_u32);
856 IMPL_LOAD4_TRANSPOSED(8, uint16_t, vld4q_u16);
857 IMPL_LOAD4_TRANSPOSED(16, uint8_t, vld4q_u8);
858 IMPL_LOAD4_TRANSPOSED(4, int32_t, vld4q_s32);
859 IMPL_LOAD4_TRANSPOSED(8, int16_t, vld4q_s16);
860 IMPL_LOAD4_TRANSPOSED(16, int8_t, vld4q_s8);
861 IMPL_LOAD4_TRANSPOSED(4, float, vld4q_f32);
862 #undef IMPL_LOAD4_TRANSPOSED
863 #elif defined(__SSE__)
864 SI void strided_load4(const float* v,
865                       Vec<4,float>& a,
866                       Vec<4,float>& b,
867                       Vec<4,float>& c,
868                       Vec<4,float>& d) {
869     using skvx::bit_pun;
870     __m128 a_ = _mm_loadu_ps(v);
871     __m128 b_ = _mm_loadu_ps(v+4);
872     __m128 c_ = _mm_loadu_ps(v+8);
873     __m128 d_ = _mm_loadu_ps(v+12);
874     _MM_TRANSPOSE4_PS(a_, b_, c_, d_);
875     a = bit_pun<Vec<4,float>>(a_);
876     b = bit_pun<Vec<4,float>>(b_);
877     c = bit_pun<Vec<4,float>>(c_);
878     d = bit_pun<Vec<4,float>>(d_);
879 }
880 #endif
881 #endif
882 
883 // De-interleaving load of 2 vectors.
884 //
885 // WARNING: These are really only supported well on NEON. Consider restructuring your data before
886 // resorting to these methods.
887 SIT void strided_load2(const T* v, skvx::Vec<1,T>& a, skvx::Vec<1,T>& b) {
888     a.val = v[0];
889     b.val = v[1];
890 }
891 SINT void strided_load2(const T* v, skvx::Vec<N,T>& a, skvx::Vec<N,T>& b) {
892     strided_load2(v, a.lo, b.lo);
893     strided_load2(v + 2*(N/2), a.hi, b.hi);
894 }
895 #if !defined(SKNX_NO_SIMD)
896 #if defined(__ARM_NEON)
897 #define IMPL_LOAD2_TRANSPOSED(N, T, VLD) \
898 SI void strided_load2(const T* v, skvx::Vec<N,T>& a, skvx::Vec<N,T>& b) { \
899     auto mat = VLD(v); \
900     a = skvx::bit_pun<skvx::Vec<N,T>>(mat.val[0]); \
901     b = skvx::bit_pun<skvx::Vec<N,T>>(mat.val[1]); \
902 }
903 IMPL_LOAD2_TRANSPOSED(2, uint32_t, vld2_u32);
904 IMPL_LOAD2_TRANSPOSED(4, uint16_t, vld2_u16);
905 IMPL_LOAD2_TRANSPOSED(8, uint8_t, vld2_u8);
906 IMPL_LOAD2_TRANSPOSED(2, int32_t, vld2_s32);
907 IMPL_LOAD2_TRANSPOSED(4, int16_t, vld2_s16);
908 IMPL_LOAD2_TRANSPOSED(8, int8_t, vld2_s8);
909 IMPL_LOAD2_TRANSPOSED(2, float, vld2_f32);
910 IMPL_LOAD2_TRANSPOSED(4, uint32_t, vld2q_u32);
911 IMPL_LOAD2_TRANSPOSED(8, uint16_t, vld2q_u16);
912 IMPL_LOAD2_TRANSPOSED(16, uint8_t, vld2q_u8);
913 IMPL_LOAD2_TRANSPOSED(4, int32_t, vld2q_s32);
914 IMPL_LOAD2_TRANSPOSED(8, int16_t, vld2q_s16);
915 IMPL_LOAD2_TRANSPOSED(16, int8_t, vld2q_s8);
916 IMPL_LOAD2_TRANSPOSED(4, float, vld2q_f32);
917 #undef IMPL_LOAD2_TRANSPOSED
918 #endif
919 #endif
920 
921 }  // namespace skvx
922 
923 #undef SINTU
924 #undef SINT
925 #undef SIN
926 #undef SIT
927 #undef SI
928 #undef SKVX_ALWAYS_INLINE
929 
930 #endif//SKVX_DEFINED
931